A restaurant's rating is usually the average of everyone who reviewed it. But a place that serves genuinely fiery food gets dragged down by people who don't like heat, and lifted by people who do. The average tells you what the crowd thought. It doesn't tell you what you'd think.
The quiz asks six things: how much heat you want, how sweet you like food, how salty, how rich, how adventurous you are, and whether you prefer dishes adapted to local tastes or made the traditional way. Each is a number from 1 to 10. Together they describe your palate as six coordinates.
When you look at a restaurant, we compare your six numbers against the six numbers of everyone who reviewed it. Reviewers whose palate sits close to yours count heavily. Reviewers far from you count very little. The score you see is the weighted result — the same reviews, read through your preferences.
A weighted score needs several reviews to mean anything. Restaurants with fewer than three show 'Not enough reviews yet' rather than a number we can't stand behind.
Not as well as the description above implies. We measured it: two readers at opposite ends of the spice scale get rankings that correlate at 0.83, and each of them correlates more closely with the plain unweighted average than with the other. The weighting reorders the list; it does not transform it. ANALYSIS.md in the repository has the numbers and the reasons.
Most of them are synthetic. 542 of the reviews in this app were generated by a script to test the matching algorithm, and they are labelled 'demo data' wherever they appear. Only a handful were written by people.
We store your email, your six taste numbers, and any reviews you write. Reviews are shown without your account attached to them. They are not fully anonymous, though — because matching needs each reviewer's taste numbers, reviews sharing an identical taste profile can be recognised as coming from the same person.